An enterprise attribution method and attribution system for industrial nodes

Abstract

The present invention belongs to the field of data mining technology, and in particular, relates to a method and system for attributing enterprises to industrial nodes. A method for attributing enterprises to industrial nodes includes: S1, constructing a target industrial chain; S2, after preprocessing the captured patent text, removing the patent text whose similarity with the target industrial chain is below a first threshold, and sending the patent text whose similarity with the target industrial chain exceeds the first threshold into a trained attribution module; S3, the attribution module attributes the applicant enterprise of the patent text to the corresponding industrial node in the target industrial chain. The present invention can quickly and accurately attribute enterprises to the corresponding industrial nodes in the target industrial chain.

Description

Technical Field

[0001] The present invention belongs to the technical field of data mining, and in particular relates to an enterprise attribution method and an attribution system for industrial nodes. Background Art

[0002] With the advancement of global industrial division of labor, the industrial chains in various fields are becoming more and more perfect, and the industrial nodes in each industrial chain are becoming more and more clear. We increasingly rely on analyzing the industrial chain containing enterprise information (i.e. enterprise name) to develop new markets, or conduct market analysis on the enterprise itself to improve its competitiveness.

[0003] At present, the process of assigning enterprises to corresponding industrial nodes in the industrial chain is mainly through manual mining of news data, and the enterprises in the news data that match the industrial node information are assigned to the corresponding industrial nodes.

[0004] However, using manual mining of news data to attribute enterprises to the corresponding industrial nodes in the target industrial chain is not only costly, but also has poor timeliness. In addition, since attributing enterprises to the corresponding target industrial nodes requires manual judgment, it is highly subjective, and different people may attribute enterprises to different industrial nodes. Therefore, the existing technology cannot quickly attribute enterprises to the corresponding industrial nodes in the target industrial chain while maintaining high accuracy. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art and to provide a method for enterprise attribution of industrial nodes, which can quickly and accurately attribute enterprises to corresponding industrial nodes in the target industrial chain.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for enterprise attribution of an industrial node comprises the following steps:

[0008] S1, build the target industrial chain;

[0009] S2, after preprocessing the captured patent texts, the patent texts whose similarity with the target industry chain is below the first threshold are eliminated, and the patent texts whose similarity with the target industry chain exceeds the first threshold are sent to the trained attribution module;

[0010] S3, the attribution module attributes the applicant enterprise of the patent text to the corresponding industrial node in the target industrial chain.

[0011] Preferably, S2 specifically further includes the following sub-steps:

[0012] S21, after crawling the patent text, it is deduplicated and cleaned to obtain patent data;

[0013] S22, constructing a first word segmentation dictionary of the target industrial chain, and segmenting the patent titles in the patent data according to the first word segmentation dictionary to obtain corresponding core contents;

[0014] S23, after calculating the similarity between each core content and the target industrial chain, the patent data corresponding to the core content whose similarity with the target industrial chain is below a first threshold is eliminated.

[0015] Preferably, the patent data obtained in S21 includes the patent name T of patent text i i and Instruction Manual Summary A i ;

[0016] S22 also includes the following sub-steps:

[0017] S221, after manually segmenting the names of each industry node, the segmentation results are added as standard segmentation words to the first segmentation dictionary D v In the example above, the part of speech of the standard participle is marked as noun;

[0018] S222, based on the first word segmentation dictionary D v Patent Name T i After segmentation, retain the patent name T i The noun in is taken as the segmentation result;

[0019] S223, remove patent name T i When the word frequency in the segmentation result is higher than the second threshold C2 and it is a noun of irrelevant vocabulary, the patent name T i The remaining nouns in the segmentation results are concatenated to obtain the corresponding core content S i : The second threshold

[0020] Among them, δ is the first parameter, Wmax(Ti) represents the patent name T i The highest frequency of nouns in the segmentation results. Wave(Ti) represents the patent name T i The average frequency of all nouns in the segmentation results. The frequency of occurrence indicates the patent name T i The noun in the segmentation result appears in the corresponding instruction manual abstract A i The frequency of irrelevant words in the text includes: method, system, and device.

[0021] Preferably, S23 also includes the following sub-steps:

[0022] S231, calculating the first similarity score between each core content and each industry node, and obtaining a first similarity score set between each core content and each industry node:

[0023]

[0024] Among them, Y(S i ,V k ) indicates the core content S i With industry node V k The first similarity score between |S i ∩V k | indicates core contentS i With industry node V k The number of identical characters in the name, |S i ∪V k | indicates core contentS i With industry node V k The total number of characters in the name, Y(S i ) indicates the core content S i The first similarity score set with each industry node; target industry chain V = {V1, V2, ..., V k ,...,V p}, where V is the target industry chain, V1 represents the most upstream industry node in the target industry chain V, that is, the first industry node in the target industry chain V, and V p represents the most downstream industrial node in the target industrial chain V, that is, the pth industrial node in the target industrial chain V, and also the last industrial node, V k represents the kth industrial node in the target industrial chain V, 1≤k≤p, and both k and p are positive integers;

[0025] S232, record Y(S i ) max is the first similarity score set Y(S i ), when Y(S i ) max When the first threshold C1 is below, the core content S is removed. i The corresponding patent data, on the contrary, will be related to the core content S i The corresponding patent data is sent to the trained attribution module;

[0026] S3 also includes the following contents: the attribution module outputs the industry node corresponding to the current patent data. At the same time, the attribution module retrieves the patent text corresponding to the current patent data in the local database, retrieves the information of the applicant company from the patent text, and then binds the applicant company to the industry node currently output by the attribution module and outputs it to the technical personnel.

[0027] Preferably, training the attribution module further includes the following steps:

[0028] Step 1: Build a sample industrial chain;

[0029] Step 2: After preprocessing the captured patent texts, the patent texts whose similarity with the sample industry chain is below the first threshold are eliminated, and the m patent texts whose similarity with the sample industry chain exceeds the first threshold are retained, which are recorded as m sample patents;

[0030] Step 3, based on m sample patents, construct a first training set, a second training set and a third training set; wherein the first training set includes m sample patents and the industry nodes to which each sample patent belongs, the second training set includes m sample patents and g industry nodes related to each sample patent, wherein m is a positive integer and g is an integer greater than 1, and the third training set includes m sample patents;

[0031] Step 4: the untrained attribution module includes a fine-tuning model, and the fine-tuning model is trained and optimized using the first training set to obtain a first model;

[0032] Step 5, in the attribution module, a reward model is obtained based on the first model copy, and the reward model is trained and optimized using the second training set to obtain a second model;

[0033] Step 6: Use the second model and the third training set to train the first model and optimize it into a third model. At this time, the attribution module training is completed.

[0034] Preferably, the specific steps of obtaining m sample patents in step 2 are the same as those in S2;

[0035] In step 3, constructing the first training set also includes steps 31 and 32:

[0036] Step 31, the sample patent data of the jth sample patent among m sample patents is recorded as P(j), 1≤j≤m, and j is an integer, and Y P(j) represents the first similarity score set between the sample patent data P(j) and each industrial node in the sample industrial chain, Y P(j) , max is the first similarity score set Y P(j) The largest first similarity score among the samples, P(j) and Y P(j) ,The industry node corresponding to max is used as the sample in the initial training set;

[0037] Step 32, manually review the correctness of the samples in the initial training set, and record the reviewed samples as the first training set TR1;

[0038] Constructing the second training set also includes steps 31′ to 32′:

[0039] Step 31′, first use the bidirectional maximum matching method based on the second word segmentation dictionary D UThe current sample patent data and each industry node are segmented, and then the jieba word segmentation is used based on the Xinhua Dictionary to segment the current sample patent data and each industry node. The obtained word segmentation is represented as one-hot and then added to the third dictionary D dict ;

[0040] Step 32′, using the Word2Vec model to obtain the word feature vectors of the current sample patent data and the word feature vectors of each industry node;

[0041] Step 33′, based on the word feature vectors of the current sample patent data and each industry node, calculate the second similarity score between the current sample patent data and each industry node and arrange them in descending order, and then use the industry nodes corresponding to the first g second similarity scores and the current sample patent data as samples of the second training set;

[0042] Step 32' also includes the following sub-steps:

[0043] Step 321′, set the context window size to o, and sequentially use each word in the current sample patent data as the central word, and obtain the context word set corresponding to each central word:

[0044] Cen(w t )={w t-o ,...,w t-q ,...,w t-1 ,w t+1 ...,w t+o}

[0045] Among them, w t represents the tth word in the current sample patent data, t and o are both positive integers greater than 1, Cen(w t ) represents the participle w t The set of context words when it is the central word, w t-o represents the (to)th word in the current sample patent data, w t+o represents the (t+o)th word in the current sample patent data, w t-q represents the (tq)th word segment in the current sample patent data, 1≤q≤o and q is an integer. When tq=0, Cen(w t )={w t-q+1 ,...,w t-1 ,w t+ 1...,w t+o}.

[0046] Step 322′, using the Word2Vec model based on the third dictionary D dictThe index position of each central word found in the search, the index position of the context word set, and the index position of the word segments contained in each industry node are used to obtain the corresponding central word feature vector, context word set feature vector, and each industry node feature vector;

[0047] Note the current sample patent data center word w t The central word feature vector is Z(w t ), the central word w t The feature vector of the corresponding context word set is {Z(w t-o ),...,Z(w t-q ),...,Z(w t-1 ),Z(w t+1 )...,Z(w t+o )}, industrial node U in sample industrial chain U a The eigenvector of is Z(U a ), Z(U a ) is the industrial node U a The sum of all word segmentation feature vectors in industry node U a The word segmentation of is obtained by step 31′;

[0048] Step 323′, obtain the characteristic vector Z[P(j)] of the current sample patent data P(j): Z[P(j)]=∑ t Z(W t ), use cosine similarity to calculate the second similarity score between the feature vector Z[P(j)] and the feature vectors of each industrial node in the sample industrial chain U; ​​take the first g industrial nodes when the second similarity score is arranged in descending order, and use them together with the current sample patent data as the samples in the second training set TR2, where g is a positive integer less than the number of nodes in the sample industrial chain.

[0049] Preferably, in step 4, the following sub-steps are also included:

[0050] Step 41, by respectively adding a first dimension reduction matrix A1 initialized by Gaussian distribution and a first dimension increase matrix B1 initialized by zero next to the attention network structure in an unsupervised LLM large language model, a fine-tuning model is obtained, and the attention network parameters in the unsupervised LLM large language model are recorded as θ0;

[0051] Step 42, after converting the first sample in the first training set TR1 into the first prompt text, after training the fine-tuning model, calculate the first objective function objective1 after training using the current sample:

[0052] The content of the first prompt text is {"instruction":"To which industry node does the knowledge involved in the following sample patent data belong?","input":"sample patent data","target":"industry node name"};

[0053]

[0054] Among them, M represents the text content corresponding to the current target, x represents the concatenation of the text content in the current instruction and input, |M| represents the length of the text content corresponding to the current target, and M f Indicates the fth word in the text content corresponding to the current target, M <f Indicates that the index position of the current target in the text content is in M f All the words before; p LLM (M f |x,M <f ) means that only the LLM large language model in the current fine-tuning model is used to predict the probability of the fth word based on the first (f-1) words in the text content corresponding to the current target; p(M f |x,M <f Indicates using the current fine-tuned model to predict the probability of the fth word based on the first (f-1) words in the text content corresponding to the current target;

[0055] Step 43, take the first dimension reduction matrix copy A1′ and the first dimension increase matrix copy B1′ corresponding to the maximum value of the first objective function objective1 gradient rise, to iteratively update the corresponding first dimension reduction matrix A1 and the first dimension increase matrix B1 in the fine-tuning model, and at the same time update the attention network parameters in the LLM large language model in the fine-tuning model to obtain a new fine-tuning model: θ′=θ+A1′·B1′, where θ′ represents the updated attention network parameters, θ represents the attention network parameters before this update, and A1′·B1′ represents the dot product of A1′ and B1′;

[0056] Step 44, after converting the next sample in the first training set TR1 into the first prompt text, the new fine-tuning model is trained, and steps 42 to 43 are repeated until the last sample in the first training set TR1 is used to train and update the fine-tuning model, and the fine-tuning model obtained at this time is recorded as the first model.

[0057] Preferably, in step 5, the following sub-steps are also included:

[0058] Step 51, after removing the sfotmax function of the output layer in the first model copy, a reward model is obtained, and the reward model outputs a scalar, i.e., a score, according to the input;

[0059] Step 52, after converting each sample in the second training set TR2 into the second prompt text, the text contents of instruction, input and target are concatenated as input to train the reward model, g = 2, and the reward model outputs the corresponding score pair, i.e., r(e, y1) and r(e, y2),

[0060] The content of the second prompt text is {"instruction":"Which industry nodes does the knowledge involved in the following sample patent data belong to?","input":"sample patent data","target":["industry node name 1","industry node name 2"]}, where the industry nodes in ["industry node 1","industry node name 2"] are arranged in descending order according to the second similarity scores calculated in step 323'.

[0061] e represents the concatenation of the text contents of instruction and input in the current second prompt text, y1 represents the text content of industry node name 1 in the current target, y2 represents the text content of industry node name 2 in the current target, r(e,y1) represents the score corresponding to y1 obtained based on e using the current reward model, and r(e,y2) represents the score corresponding to y2 obtained based on e using the current reward model;

[0062] Step 53, after each Q1 sample is used to train the reward model, the second objective function objective2 is calculated once, and the reward model is optimized in the direction of the gradient descent of the second objective function objective2:

[0063]

[0064] Among them, (e, y1, y2)~TR2 represent the samples in the second training set TR2 represents the expected value of the current sample (e, y1, y2), σ[·] represents the sigmoid function, 0<Q1≤m, and Q1 is a positive integer;

[0065] Step 54, the reward model after repeating steps 52 to 53 for R1 training rounds is recorded as the second model. One training round means training the reward model once using all samples in the second training set TR2.

[0066] Preferably, in step 6, the following sub-steps are also included:

[0067] Step 61, after converting each sample in the third training set TR3 into a third prompt text, the text contents of instruction and input in the third prompt text are concatenated as input to train the first model, and the first model outputs the corresponding answer, that is, the name of an industry node;

[0068] The content of the third prompt text is {“instruction”: “To which industry node does the knowledge involved in the following sample patent data belong?”, “input”: “sample patent data”};

[0069] Step 62, after each Q2 samples are used to train the first model, the current sample and the corresponding answer are converted into the first prompt text, and the text contents of instruction, input and target in the current first prompt text are concatenated as the input of the second model, and the output r(c,M) corresponding to the second model is obtained, wherein c represents the concatenation of the text contents of instruction and input in the current first prompt text, M represents the text content of target in the current first prompt text, i.e., the name of the industry node, r(c,M) represents the score corresponding to M obtained based on c using the current second model, 0<Q2≤m, and Q1 is a positive integer;

[0070] Calculate the third objective function objective3 once, and optimize the first model in the direction of the gradient ascent of the third objective function objective3:

[0071] Among them, β is a constant parameter, (c)~TR3 means that the text content of the input in c is a sample in the third training set TR3, represents the expected value under the current c, π SFT (M|c) represents the probability that the initial first model obtains output M when the input is c, π RL (M|c) represents the probability that the current first model obtains the output M when the input is c, and the initial first model refers to the first model obtained in step 44;

[0072] Step 63, after repeating steps 61 to 62 for R2 training rounds, the first model is recorded as the third model. One training round means training the first model once using all samples in the third training set TR3, and the attribution module training is completed at this time.

[0073] The present invention also provides an enterprise attribution system for industrial nodes, including: a capture module, a preprocessing module, an attribution module, and an output module. The capture module is used to capture patent texts and send them to the preprocessing module; the preprocessing module is used to preprocess the patent texts, and send the patent texts whose similarity with the target industrial chain exceeds a first threshold after preprocessing to the attribution module; the attribution module is used to attribute the applicant enterprise of the patent text to the corresponding industrial node in the target industrial chain, and send the attribution result to the output module; the output module is used to output the attribution result to the technician; each module is programmed or configured to execute the steps of an enterprise attribution method for an industrial node as described above.

[0074] The beneficial effects of the present invention are:

[0075] (1) The enterprise attribution method of an industrial node of the present invention only needs to spend a certain amount of time in the early stage to train the attribution module. After the technicians have constructed the target industrial chain, the enterprises can be accurately, efficiently and comprehensively attributed to the corresponding industrial nodes in the target industrial chain. The whole process does not require the technicians to understand the enterprise information at all, because the present invention attributes the enterprise to the industrial node based on the patent text. The patent text is strongly related to the main business and core competitiveness of an enterprise. The existence of the patent text proves that the enterprise has potential or has already had a large production in the industrial node to which the patent text belongs. Compared with the prior art that attributes enterprises through news data containing a lot of noise, the enterprise attribution method of the present invention is based on highly technical patent texts for data mining, which is more accurate and scientific. Due to the language style, grammatical errors, spelling errors, repeated content, false reports and other problems of news reports, a lot of noise may be introduced into the news data, resulting in the inability to determine the truly useful information in the news data, thereby causing the enterprise to be attributed to the wrong industrial node.

[0076] (2) The present invention directly removes patent texts whose core content has a maximum first similarity score below a first threshold with any industrial node in the target industrial chain. That is, the present invention directly removes patent texts that are basically irrelevant to the target industrial chain, and only retains patent data related to the target industrial chain and sends it to the subsequently trained attribution module, thereby reducing the time consumption caused by irrelevant patent texts, greatly improving the work efficiency of the subsequent attribution module, and also improving the time consumption of the entire process of attributing enterprises to industrial nodes.

[0077] (3) The present invention is a method for attributing enterprises based on different patent texts. Therefore, when an enterprise applies for multiple patents involving different industrial nodes, the present invention will attribute the enterprise to different industrial nodes based on the specific patent texts, so that technical personnel can clearly see the multiple industrial nodes involved by the enterprise in subsequent market analysis.

[0078] (4) When the present invention trains the home module:

[0079] ① The present invention uses the retained patent data as sample patent data after filtering out patent data that is basically irrelevant to the sample industrial chain, and forms an initial training set with the sample patent data and the industrial node with the largest first similarity score. Manual verification of the correctness of the samples in the initial training set is sufficient to obtain an accurate first training set. Therefore, the process of constructing the first training set in the present invention not only reduces the workload of personnel and ensures the accuracy of the samples in the first training set, but also greatly shortens the time to obtain the first training set with the same number of samples.

[0080] ② When constructing the second training set, the Word2Vec model is used. Based on the connection between the front and back word segmentations (i.e., context) in a sample patent data, a sample patent data is converted into several central word feature vectors, and then the feature vector of the current sample patent data is obtained based on the central word feature vector. The feature vector of the industrial node in the sample industrial chain is also obtained in this way. Finally, the similarity score between the current sample patent data and the industrial node is calculated based on the feature vector. Therefore, the samples of the second training set are from a different direction from the first training set, and the g industrial nodes with the highest matching degree with the current sample patent data are obtained; and in the process of obtaining the second training set, the Word2Vec model used to obtain the feature vector of the next sample patent data is optimized based on the feature vector obtained from the previous sample patent data. Therefore, the process of constructing the second training set is accompanied by the continuous optimization of the Word2Vec model, which not only shortens the time to construct the second training set, but also improves the accuracy of the samples in the second training set.

[0081] ③ The first model obtained after training and optimization of the first training set in the present invention can accurately calculate the similarity between the current patent data and each industrial node in the target industrial chain, and take the industrial node with the highest similarity as the output; and the second model obtained after training and optimization of the second training set in the present invention can further amplify the difference in the similarity between the patent data and each industrial node.

[0082] ④ The attribution module of the present invention creatively performs composite training during the training process, that is, the first model is trained and optimized into the third model using the second model and the third training set. This will make the third model more different in calculating the similarity scores between the same patent data and different industrial nodes, and will not have the same similarity scores between the same patent data and different industrial nodes. This also allows the trained attribution module to output the industrial node to which it belongs more quickly, accurately and scientifically based on a patent data.

[0083] From the above analysis, it can be seen that when the present invention trains the attribution module, whether it is the construction of the training set or the training process, there is no need for manual large-scale data labeling and supervision, so the attribution module can be quickly trained, so that the trained attribution module can accurately attribute the patent application enterprise to the corresponding industry node.

[0084] (5) Compared with manual methods, the enterprise attribution method of the present invention is not only efficient and accurate, but also saves a lot of labor costs. In addition, the standard for attributing enterprises to corresponding industrial nodes is unified, which is not as subjective as manual judgment, and naturally avoids the situation where the accuracy rate fluctuates. In addition, compared with the prior art of directly training an LLM large language model and using the trained LLM large language model to attribute enterprises to industrial nodes, the trained attribution module of the present invention has a higher accuracy rate in attributing enterprises to industrial nodes.

[0085] (6) Through the present invention, technical personnel can quickly and accurately obtain the target industrial chain containing the enterprise name in each industrial node to conduct new market analysis, analyze the core competitiveness of the enterprise itself, understand the industrial development trend, and thus adjust the industrial structure and achieve industrial upgrading; understand the development status of upstream and downstream industries, strengthen cooperation with upstream and downstream enterprises, achieve coordinated development of the industrial chain, and improve the competitiveness of the entire industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A flowchart of a method for enterprise attribution of an industry node;

[0087] Figure 2 A time-consuming comparison curve diagram between the enterprise attribution method of the present invention and the prior art;

[0088] Figure 3 It is a curve chart comparing the accuracy between the enterprise attribution method of the present invention and the prior art. DETAILED DESCRIPTION

[0089] In order to make the technical solution of the present invention clearer and more specific, the present invention is clearly and completely described below in conjunction with the accompanying drawings. Any equivalent replacement of the technical features of the technical solution of the present invention and any solution derived by conventional reasoning by ordinary technicians in the field without making any creative work shall fall within the protection scope of the present invention.

[0090] Because the present invention is aimed at enterprises rather than individuals, for the convenience of description, the case where an enterprise is the applicant is defined as an applicant enterprise in the present invention.

[0091] A flowchart of a method for enterprise attribution of industrial nodes is shown below: Figure 1 As shown, the following steps are included:

[0092] S1, build the target industrial chain;

[0093] S2, after preprocessing the captured patent texts, the patent texts whose similarity with the target industry chain is below the first threshold are eliminated, and the patent texts whose similarity with the target industry chain exceeds the first threshold are sent to the trained attribution module;

[0094] S3, the attribution module attributes the applicant enterprise of the patent text to the corresponding industrial node in the target industrial chain.

[0095] In S1, the following are also included:

[0096] Build the target industry chain V = {V1, V2, ..., V k ,...,V p}, where V is the target industry chain, V1 represents the most upstream industry node in the target industry chain V, that is, the first industry node in the target industry chain V, and V p represents the most downstream industrial node in the target industrial chain V, that is, the pth industrial node in the target industrial chain V, and also the last industrial node, V k Represents the kth industrial node in the target industrial chain V, 1≤k≤p, and k and p are both positive integers.

[0097] The technical personnel name each industrial node in the target industrial chain V according to the actual needs of the market analysis and the industrial standards of the industry to which it belongs.

[0098] S2 also includes the following sub-steps:

[0099] S21, after crawling the patent text, it is deduplicated and cleaned to obtain patent data;

[0100] S22, constructing a first word segmentation dictionary of the target industrial chain, and segmenting the patent titles in the patent data according to the first word segmentation dictionary to obtain corresponding core contents;

[0101] S23, after calculating the similarity between each core content and the target industrial chain, the patent data corresponding to the core content whose similarity with the target industrial chain is below a first threshold is eliminated.

[0102] In S21, deduplication and cleaning are preprocessing.

[0103] S21 also includes the following sub-steps:

[0104] S211, capture patent text from the patent database. If the captured patent text is in a foreign language, translate the captured patent text into Chinese.

[0105] In S211, the captured patent text type is invention or utility model.

[0106] In this embodiment, the patent database from which the patent text is captured includes but is not limited to the database of the State Intellectual Property Office of China.

[0107] S212, compare the captured patent text with the patent text stored in the local database. If the captured patent text is repeated with the patent text stored in the local database, the captured patent text is discarded, otherwise the captured patent text is stored in the local database.

[0108] The captured patent text is duplicated with the patent text stored in the local database, which means any of the following situations occurs:

[0109] ① The content of the currently captured patent text is exactly the same as the patent text stored in the local database (claims and description);

[0110] ②The patent text currently captured is the authorization announcement text of the invention patent text stored in the local database.

[0111] In the present invention, each time a new target industry chain is constructed, a local database corresponding to the current new target industry chain is established. In S212, the captured patent texts are deduplicated to ensure that there is only one duplicate patent text in the local database, thereby avoiding the subsequent repeated patents from being calculated and analyzed multiple times, and reducing the computing overhead and memory resource usage.

[0112] S213, after cleaning the newly added patent texts in the local database and removing illegal characters, only the patent names and description abstracts in each patent text are retained, that is, the patent name T of patent text i is i and Instruction Manual Summary A i As the corresponding patent data.

[0113] Illegal characters refer to characters such as spaces, underscores, backslashes, etc. that are not Arabic numerals and do not meet the rules for using Chinese characters and Chinese article symbols, such as 、 , <italic> 、< / italic> wait.

[0114] S22 also includes the following sub-steps:

[0115] S221, after manually segmenting the names of each industry node, the segmentation results are added as standard segmentation words to the first segmentation dictionary D v In the example above, the part of speech of the standard participle is marked as noun.

[0116] S222, based on the first word segmentation dictionary D v Patent Name T i After segmentation, retain the patent name T i The nouns in are taken as the segmentation results.

[0117] S223, remove patent name T i When the word frequency in the segmentation result is higher than the second threshold C2 and it is a noun of irrelevant vocabulary, the patent name T i The remaining nouns in the segmentation results are concatenated to obtain the corresponding core content S i : The second threshold

[0118] Wherein, δ is the first parameter, in this embodiment, δ=0.75, Wmax(Ti) represents the patent name T i The highest frequency of nouns in the segmentation results. Wave(Ti) represents the patent name T i The average frequency of all nouns in the segmentation results. The frequency of occurrence indicates the patent name T i The noun in the segmentation result appears in the corresponding instruction manual abstract A i The frequency of irrelevant words in the text includes: method, system, and device.

[0119] Because the core of a patent text may be just a manufacturing method, but the patent may contain a "system" or "device" that applies the manufacturing method without any substantial content, which may increase the frequency of "system" or "device" in the patent's abstract. Therefore, when irrelevant words appear in the segmentation of the patent name and the frequency of the irrelevant words in the corresponding abstract is also high, in order to avoid introducing noise into the core content, we will remove such irrelevant words that appear too frequently in S223.

[0120] S222 also includes sub-steps S222a to S222c:

[0121] S222a, using the bidirectional maximum matching method based on the first word segmentation dictionary D v Patent Name T i The patent name T i All standard participles separated from the middle segmentation are marked as noun parts of speech;

[0122] S222b, using Jieba segmentation based on Xinhua Dictionary to identify patent names T i The rest of the words except the standard participles are segmented and the real parts of speech are marked;

[0123] S222c, reserved patent name T i The segmented noun is used as the patent name T i The segmentation result.

[0124] In S222a, the first word segmentation dictionary D vThe true part-of-speech of the internal standard segmentation may include verbs, such as "manufacture". However, to reduce the noise introduced to the subsequent core content, we also mark "manufacture" as a noun in the standard segmentation. Therefore, as can be seen in S222a, for the patent name T i in which as long as the words are standard segmentations, regardless of their true part-of-speech, they will be marked as nouns; while in S222b, for the remaining part of the patent name T i the rest is segmented using jieba based on the Xinhua Dictionary and the true part-of-speech of the segmented words is marked.

[0125] In S222a to S222c, by performing two segmentation operations with different priorities (the first priority is S222a and the second priority is S222b) on the patent name T i in sequence, the segmentation result of the patent name T i is obtained. This is to ensure that when segmenting the patent name T i the standard segmentation will not be split. For example, if the standard segmentation is "new energy", then we can accurately segment out "new energy" instead of splitting it into "new" and "energy".

[0126] In S23, the following sub-steps are also included:

[0127] S231, calculate the first similarity score between each core content and each industrial node, and obtain the set of the first similarity scores between each core content and each industrial node:

[0128]

[0129] Among them, Y(S i , V k ) represents the first similarity score between the core content S i and the industrial node V k . |S i ∩V k | represents the number of identical characters in the names of the core content S i and the industrial node V k . |S i ∪V k | represents the total number of characters in the names of the core content S i and the industrial node V k . Y(S i ) represents the set of the first similarity scores between the core content S i and each industrial node.

[0130] S232, denote Y(S i ) max as the set of the first similarity scores Y(S i), when Y(S i ) max When the first threshold C1 is below, the core content S is removed. i The corresponding patent data, on the contrary, will be related to the core content S i The corresponding patent data is sent to the trained attribution module.

[0131] In this embodiment, the first threshold C1 is 0.25.

[0132] When the maximum first similarity score between the core content and any industrial node in the target industrial chain is below the first threshold C1, the patent text corresponding to the core content is basically irrelevant to the target industrial chain, and its patent data does not need to be further processed, so it is eliminated. The present invention screens out a large part of the patent texts irrelevant to the target industrial chain through S231-S232, and only retains the patent data related to the target industrial chain and sends it to the subsequently trained attribution module, which greatly improves the work efficiency of the subsequent attribution module and reduces the time consumption caused by irrelevant patent texts.

[0133] Training the attribution module includes the following steps:

[0134] Step 1: Build a sample industrial chain;

[0135] Step 2: After preprocessing the captured patent texts, the patent texts whose similarity with the sample industry chain is below the first threshold are eliminated, and the m patent texts whose similarity with the sample industry chain exceeds the first threshold are retained, which are recorded as m sample patents;

[0136] Step 3, based on m sample patents, construct a first training set, a second training set and a third training set; wherein the first training set includes m sample patents and the industry nodes to which each sample patent belongs, the second training set includes m sample patents and g industry nodes related to each sample patent, wherein m is a positive integer and g is an integer greater than 1, and the third training set includes m sample patents;

[0137] Step 4: the untrained attribution module includes a fine-tuning model, and the fine-tuning model is trained and optimized using the first training set to obtain a first model;

[0138] Step 5, in the attribution module, a reward model is obtained based on the first model copy, and the reward model is trained and optimized using the second training set to obtain a second model;

[0139] Step 6: Use the second model and the third training set to train the first model and optimize it into a third model. At this time, the attribution module training is completed.

[0140] The specific content of step 1 is similar to that of S1, and the sub-steps included in step 2 are also similar to S2 and its sub-steps, which will not be repeated here.

[0141] In step 3, constructing the first training set also includes steps 31 and 32:

[0142] Step 31, the sample patent data of the jth sample patent among m sample patents is recorded as P(j), 1≤j≤m, and j is an integer, and Y P(j) represents the first similarity score set between the sample patent data P(j) and each industrial node in the sample industrial chain, Y P(j) , max is the first similarity score set Y P(j) The largest first similarity score among the samples, P(j) and Y P(j) , the industry node corresponding to max is used as the sample in the initial training set.

[0143] Step 32, manually review the correctness of the samples in the initial training set, and record the reviewed samples as the first training set TR1.

[0144] If the sample patent data P(j) truly belongs to Y P(j) ,max corresponds to the industry node, the sample is correct, and the sample remains unchanged after manual review; if the sample patent data P(j) does not really belong to Y P(j) ,max corresponds to the industry node, the sample is wrong, and after manual review, the industry node in the wrong sample will be modified to the industry node to which the sample patent data P(j) actually belongs. For example: in the initial training set, the sample patent data P(j) is bound to the industry node V5 as an input and output sample, but after manual review, it is found that the sample patent data P(j) actually belongs to the industry node V8. In the first training set TR1, the sample patent data P(j) is bound to the industry node V8 as an input and output sample.

[0145] The present invention uses the retained patent data as sample patent data after filtering out patent data that is basically irrelevant to the sample industrial chain, and forms an initial training set with the sample patent data and the industrial node with the largest first similarity score, and only needs to manually review the correctness of the samples in the initial training set in the future to obtain an accurate first training set TR1, that is, after manually reviewing 1000 samples in the initial training set, 1000 new samples will be added to the first training set TR1; if the first training set TR1 is directly constructed manually as in the prior art, 1000 patent texts are captured, and 1000 patent data are obtained, the staff needs to spend a lot of time and energy to understand the content of each patent data, so as to bind the patent data with the corresponding industrial node into input and output samples, but in this way, there may be only 10 samples in the first training set TR1, because the remaining 990 patent data are irrelevant to any industrial node in the sample industrial chain. Therefore, the process of constructing the first training set TR1 in the present invention not only reduces the workload of personnel, ensures the accuracy of the samples in the first training set TR1, but also greatly shortens the time to obtain the first training set TR1 with the same number of samples.

[0146] In step 3, constructing the second training set further includes steps 31′ to 32′:

[0147] Step 31′, first use the bidirectional maximum matching method based on the second word segmentation dictionary D U The current sample patent data and each industry node are segmented, and then the jieba word segmentation is used based on the Xinhua Dictionary to segment the current sample patent data and each industry node. The obtained word segmentation is represented as one-hot and then added to the third dictionary D dict .

[0148] In the present invention, the second word segmentation dictionary D is obtained. U The steps are the same as S221.

[0149] Step 32 ′, using the Word2Vec model to obtain the word feature vectors of the current sample patent data and the word feature vectors of each industry node.

[0150] Step 33′, based on the word feature vectors of the current sample patent data and each industry node, calculate the second similarity score between the current sample patent data and each industry node and arrange them in descending order, and then use the industry nodes corresponding to the first g second similarity scores and the current sample patent data as samples of the second training set.

[0151] Step 32' also includes the following sub-steps:

[0152] Step 321′, set the context window size to o, and sequentially use each word in the current sample patent data as the central word, and obtain the context word set corresponding to each central word:

[0153] Cen(w t )={w t-o ,...,w t-q ,...,w t-1 ,w t+1 ...,w t+o}

[0154] Among them, w t represents the tth word in the current sample patent data, t and o are both positive integers greater than 1, Cen(w t ) represents the participle w t The set of context words when it is the central word, w t-o represents the (to)th word in the current sample patent data, w t+o represents the (t+o)th word in the current sample patent data, w t-q represents the (tq)th word segment in the current sample patent data, 1≤q≤o and q is an integer. When tq=0, Cen(w t )={w t-q+1 ,...,w t-1 ,w t+1 ...,w t+o}.

[0155] Step 322′, using the Word2Vec model based on the third dictionary D dict The index position of each central word found in the search, the index position of the context word set, and the index position of the word segments contained in each industry node are used to obtain the corresponding central word feature vector, context word set feature vector, and each industry node feature vector;

[0156] Note the current sample patent data center word w t The central word feature vector is Z(w t ), the central word w t The feature vector of the corresponding context word set is {Z(w t-o ),...,Z(w t-q ),...,Z(w t-1 ),Z(w t+1 )...,Z(w t+o )}, industrial node U in sample industrial chain U a The eigenvector of is Z(U a ), Z(U a ) is the industrial node U a The sum of all word segmentation feature vectors in industry node Ua The word segmentation of is obtained by step 31′. The feature vector is a column vector.

[0157] Step 323′, obtain the characteristic vector Z[P(j)] of the current sample patent data P(j): Z[P(j)]=Σ t Z(W t ), use cosine similarity to calculate the second similarity score between the feature vector Z[P(j)] and the feature vectors of each industrial node in the sample industrial chain U; ​​take the first g industrial nodes when the second similarity score is arranged in descending order, and use them together with the current sample patent data as the samples in the second training set TR2, where g is a positive integer less than the number of nodes in the sample industrial chain.

[0158] Using cosine similarity to calculate the similarity score is a prior art and will not be described in detail here.

[0159] In this embodiment, g is 2.

[0160] Optionally, after step 322′, step 323″ is further included:

[0161] Step 323″, calculate the first loss function Loss1, and continuously optimize the Word2Vec model in the direction of the gradient descent of the first loss function L1:

[0162]

[0163]

[0164] Where sum{·} represents the sum function, u(w t ) is an intermediate variable, [u(w t )] T represents the intermediate variable u(w t ), Z(D b ) indicates the third dictionary D dict The bth participle D in b Based on the feature vector obtained by the Word2Vec model, |D dict | indicates the current third dictionary D dict The number of words in , p(w t |w t-o ,...,w t-p ,...,w t-1 ,w t+1 ,...,w t+o ) indicates that the context word set

[0165] {w t-o ,...,w t-q ,...,w t-1 ,wt+1 ...,w t+o} to predict the central word w t , L represents the number of segmented words in the current sample patent data.

[0166] In the present invention, the Word2Vec model is used. Based on the connection between the front and back participles (i.e., context) in a piece of sample patent data, a piece of sample patent data is converted into several central word feature vectors, and then the feature vector of the current sample patent data is obtained based on the central word feature vector. The industrial nodes in the sample industrial chain also obtain feature vectors in this way. Finally, the similarity score between the current sample patent data and the industrial node is calculated based on the feature vector. Therefore, the samples of the second training set TR2 are from a different direction from the first training set TR1, and the g industrial nodes with the highest matching degree with the current sample patent data are obtained. And in the process of obtaining the second training set TR2, the Word2Vec model used to obtain the feature vector of the next sample patent data is optimized based on the feature vector obtained from the previous piece of sample patent data. Therefore, in the process of constructing the second training set TR2, the Word2Vec model is continuously optimized, which not only shortens the time of constructing the second training set TR2, but also improves the accuracy of the samples in the second training set TR2.

[0167] In step 4, the following sub-steps are also included:

[0168] Step 41, by respectively adding a first dimension reduction matrix A1 initialized by Gaussian distribution and a first dimension increase matrix B1 initialized by zero next to the attention network structure in an unsupervised LLM large language model, a fine-tuning model is obtained, and the attention network parameters in the unsupervised LLM large language model are recorded as θ0;

[0169] Step 42, after converting the first sample in the first training set TR1 into the first prompt text, after training the fine-tuning model, calculate the first objective function objective1 after training using the current sample:

[0170] The content of the first prompt text is {"instruction": "To which industry node does the knowledge involved in the following sample patent data belong?", "input": "sample patent data", "target": "industry node name"}. The first prompt text is in json format.

[0171]

[0172] Among them, M represents the text content corresponding to the current target, x represents the concatenation of the text content in the current instruction and input, |M| represents the length of the text content corresponding to the current target, and Mf Indicates the fth word in the text content corresponding to the current target, M <f Indicates that the index position of the current target in the text content is in M f All the words before; p LLM (M f |x,M <f ) means that only the LLM large language model in the current fine-tuning model is used to predict the probability of the fth word based on the first (f-1) words in the text content corresponding to the current target; p(M f |x,M <f It means using the current fine-tuned model to predict the probability of the fth word based on the first (f-1) words in the text content corresponding to the current target.

[0173] Step 43, take the first dimension reduction matrix copy A1′ and the first dimension increase matrix copy B1′ corresponding to the first objective function objective1 gradient rise and the maximum value, to iteratively update the corresponding first dimension reduction matrix A1 and the first dimension increase matrix B1 in the fine-tuning model, and at the same time update the attention network parameters in the LLM large language model in the fine-tuning model to obtain a new fine-tuning model: θ′=θ+A1′·B1′,

[0174] Among them, θ′ represents the updated attention network parameters, θ represents the attention network parameters before this update, and A1′·B1′ represents the dot product of A1′ and B1′.

[0175] Step 44, after converting the next sample in the first training set TR1 into the first prompt text, the new fine-tuning model is trained, and steps 42 to 43 are repeated until the last sample in the first training set TR1 is used to train and update the fine-tuning model, and the fine-tuning model obtained at this time is recorded as the first model.

[0176] In step 5, the following sub-steps are also included:

[0177] Step 51, after removing the sfotmax function of the output layer in the first model copy, a reward model is obtained, and the reward model outputs a scalar, i.e., a score, based on the input.

[0178] After the sfotmax function of the output layer in the first model copy is removed, it becomes a linear layer.

[0179] Step 52, after converting each sample in the second training set TR2 into the second prompt text, the text contents of instruction, input and target are concatenated as input to train the reward model, and the reward model outputs the corresponding score pair, i.e., r(e,y1) and r(e,y2).

[0180] The content of the second prompt text is {"instruction":"Which industry nodes does the knowledge involved in the following sample patent data belong to?","input":"sample patent data","target":["industry node name 1","industry node name 2"]}, where the industry nodes in ["industry node 1","industry node name 2"] are arranged in descending order according to the second similarity scores calculated in step 323', and the second prompt text is in json format.

[0181] e represents the concatenation of the text contents of instruction and input in the current second prompt text, y1 represents the text content of industry node name 1 in the current target, y2 represents the text content of industry node name 2 in the current target, r(e,y1) represents the score corresponding to y1 obtained based on e using the current reward model, and r(e,y2) represents the score corresponding to y2 obtained based on e using the current reward model.

[0182] Step 53, after each Q1 sample is used to train the reward model, the second objective function objective2 is calculated once, and the reward model is optimized in the direction of the gradient descent of the second objective function objective2:

[0183]

[0184] Among them, (e, y1, y2)~TR2 represent the samples in the second training set TR2 represents the expected value of the current sample (e, y1, y2), σ[·] represents the sigmoid function, 0<Q1≤m, and Q1 is a positive integer.

[0185] For example, the second training set TR2 contains m=1000 samples. After each time the reward model is trained using Q1=5 samples, the reward model needs to be optimized once.

[0186] Step 54, the reward model after repeating steps 52 to 53 for R1 training rounds is recorded as the second model. One training round means training the reward model once using all samples in the second training set TR2.

[0187] In this embodiment, R1=1.

[0188] Step 6 also includes the following specific sub-steps:

[0189] Step 61, after converting each sample in the third training set TR3 into a third prompt text, the text contents of instruction and input in the third prompt text are concatenated as input to train the first model, and the first model outputs the corresponding answer, i.e., the name of an industry node.

[0190] The content of the third prompt text is {"instruction":"To which industry node does the knowledge involved in the following sample patent data belong?","input":"sample patent data"}, and the third prompt text is in json format;

[0191] Step 62, after each Q2 samples are used to train the first model, the current sample and the corresponding answer are converted into the first prompt text as described in step 42, and the text contents of instruction, input and target in the current first prompt text are concatenated as the input of the second model to obtain the output r(c,M) corresponding to the second model, wherein c represents the concatenation of the text contents of instruction and input in the current first prompt text, M represents the text content of target in the current first prompt text, i.e., the name of the industry node, and r(c,M) represents the score corresponding to M obtained based on c using the current second model, 0<Q2≤m, and Q2 is a positive integer;

[0192] Calculate the third objective function objective3 once, and optimize the first model in the direction of the gradient ascent of the third objective function objective3:

[0193] Wherein, β is a constant parameter, which is 0.02 in this embodiment, (c)-TR3 means that the text content of input in c is a sample in the third training set TR3, represents the expected value under the current c, π SFT (M|c represents the probability of the initial first model obtaining output M when the input is c, that is, the product of the output layer softmax function value when the initial first model obtains output M when the input is c, π RL (M|c) represents the probability of the current first model obtaining output M when the input is c, that is, the product of the output layer softmax function value when the current first model obtains output M when the input is c. The initial first model refers to the first model obtained in step 44.

[0194] Step 63, after repeating steps 61 to 62 for R2 training rounds, the first model is recorded as the third model. One training round means training the first model once using all samples in the third training set TR3, and the attribution module training is completed at this time.

[0195] In this embodiment, R2=1.

[0196] The first model obtained after training and optimization of the first training set in the present invention can accurately calculate the similarity between the current patent data and each industrial node in the target industrial chain, and output the industrial node with the highest similarity; and the second model obtained after training and optimization of the second training set in the present invention can further amplify the difference in the similarity between the patent data and each industrial node. Therefore, in step 6, after the first model is trained and optimized using the second model and the third training set, the final third model will have a greater difference in calculating the similarity scores between the same patent data and different industrial nodes, and there will be no situation where the similarity scores between the same patent data and different industrial nodes are the same, which also allows the trained attribution module to output the industrial node to which it belongs more quickly and accurately based on a patent data.

[0197] The following are also included in S3:

[0198] The attribution module outputs the industry node corresponding to the current patent data. At the same time, the attribution module retrieves the patent text corresponding to the current patent data in the local database, retrieves the information of the applicant company from the patent text, and then binds the applicant company to the industry node currently output by the attribution module and outputs it to the technical staff.

[0199] Subsequent technical personnel can attribute each applicant enterprise to each industrial node in the target industrial chain based on the output of the attribution module to form an industrial chain information map.

[0200] The enterprise attribution method of an industrial node of the present invention only needs to spend a certain amount of time in the early stage to train the attribution module. After the technicians build the target industrial chain, the enterprises can be accurately, efficiently and comprehensively attributed to the corresponding industrial nodes in the target industrial chain. The whole process does not require the technicians to understand the enterprise information at all, because the present invention attributes the enterprise to the industrial node on the basis of the patent text. The patent text is strongly related to the main business and core competitiveness of an enterprise. The existence of the patent text proves that the enterprise has potential or has already had a large production in the industrial node to which the patent text belongs. Compared with the prior art, the enterprise attribution method of the present invention is based on the highly technical patent text for data mining, which is more accurate and scientific. Due to the language style, grammatical errors, spelling errors, repeated content, false reports and other problems of news reports, a large amount of noise may be introduced into the news data, resulting in the inability to determine the truly useful information in the news data, thereby causing the enterprise to be attributed to the wrong industrial node.

[0201] The present invention directly eliminates patent texts whose core content has a maximum first similarity score below the first threshold C1 with respect to any industrial node in the target industrial chain. That is, the present invention directly eliminates patent texts that are basically irrelevant to the target industrial chain, and only retains patent data related to the target industrial chain and sends them to the subsequently trained attribution module, thereby reducing the time consumption caused by irrelevant patent texts, greatly improving the work efficiency of the subsequent attribution module, and also improving the time consumption of the entire process of attributing enterprises to industrial nodes.

[0202] The present invention is a method for enterprise attribution based on different patent texts. Therefore, when an enterprise applies for multiple patents involving different industrial nodes, the present invention will attribute the enterprise to different industrial nodes according to the specific patent texts, so that technical personnel can clearly see the multiple industrial nodes involved by the enterprise in subsequent market analysis.

[0203] When training the attribution module:

[0204] ① The present invention uses the retained patent data as sample patent data after filtering out patent data that is basically irrelevant to the sample industrial chain, and forms an initial training set with the sample patent data and the industrial node with the largest first similarity score. Manual verification of the correctness of the samples in the initial training set is sufficient to obtain an accurate first training set TR1. Therefore, the process of constructing the first training set TR1 in the present invention not only reduces the workload of personnel and ensures the accuracy of the samples in the first training set TR1, but also greatly shortens the time to obtain the first training set TR1 with the same number of samples.

[0205] ② When constructing the second training set TR2, the Word2Vec model is used. Based on the connection between the front and back word segmentations (i.e., context) in a sample patent data, a sample patent data is converted into several central word feature vectors, and then the feature vector of the current sample patent data is obtained based on the central word feature vector. The feature vector of the industrial node in the sample industrial chain is also obtained in this way. Finally, the similarity score between the current sample patent data and the industrial node is calculated based on the feature vector. Therefore, the samples of the second training set TR2 are from a different direction from the first training set TR1, and the g industrial nodes with the highest matching degree with the current sample patent data are obtained; and in the process of obtaining the second training set TR2, the Word2Vec model used to obtain the feature vector of the next sample patent data is optimized based on the feature vector obtained from the previous sample patent data. Therefore, in the process of constructing the second training set TR2, the Word2Vec model is continuously optimized, which not only shortens the time to construct the second training set TR2, but also improves the accuracy of the samples in the second training set TR2.

[0206] ③ The first model obtained after training and optimization of the first training set in the present invention can accurately calculate the similarity between the current patent data and each industrial node in the target industrial chain, and take the industrial node with the highest similarity as the output; and the second model obtained after training and optimization of the second training set in the present invention can further amplify the difference in the similarity between the patent data and each industrial node.

[0207] ④ The attribution module of the present invention creatively performs composite training during the training process, that is, the first model is trained and optimized into the third model using the second model and the third training set. This will make the third model more different in calculating the similarity scores between the same patent data and different industrial nodes, and will not have the same similarity scores between the same patent data and different industrial nodes. This also allows the trained attribution module to output the industrial node to which it belongs more quickly, accurately and scientifically based on a patent data.

[0208] From the above analysis, it can be seen that when the present invention trains the attribution module, whether it is the construction of the training set or the training process, there is no need for manual large-scale data labeling and supervision, so the attribution module can be quickly trained, so that the trained attribution module can accurately attribute the patent application enterprise to the corresponding industry node.

[0209] Compared with manual work, the enterprise attribution method of an industrial node of the present invention is not only efficient and accurate, but also saves a lot of labor costs. In addition, the standard for attributing enterprises to corresponding industrial nodes is unified, and it is not as subjective as manual judgment, which naturally avoids the situation where the accuracy rate fluctuates. And compared with the prior art of directly training an LLM large language model and using the trained LLM large language model to attribute enterprises to industrial nodes, the trained attribution module of the present invention has a higher accuracy rate in attributing enterprises to industrial nodes.

[0210] Through the present invention, technical personnel can quickly and accurately obtain the target industrial chain containing the company names in each industrial node to conduct new market analysis, analyze the core competitiveness of the company itself, understand the industry development trend, and thus adjust the industrial structure and achieve industrial upgrading; understand the development status of upstream and downstream industries, strengthen cooperation with upstream and downstream companies, achieve coordinated development of the industrial chain, and improve the competitiveness of the entire industry.

[0211] For the same 700 patent texts and target industrial chains (some of the 700 patent texts are completely irrelevant to the target industrial chain), with high-quality, multiple-reviewed manual inspection results as the standard, the time-consuming and accuracy comparison curves between the enterprise attribution method of the present invention and the prior art are as follows: Figure 2-3As shown, there are two types of prior art. Prior art 1 is to manually attribute the applicants of 700 patent texts to the corresponding industrial nodes in the target industrial chain, and prior art 2 is to use the trained LLM large language model to attribute the applicants of 700 patent texts to the corresponding industrial nodes in the target industrial chain. It can be seen intuitively that with the increase in the number of patent texts, the accuracy of the present invention is significantly higher than that of prior art 1 and prior art 2, and the overall accuracy is relatively stable; the total time consumption of the present invention is similar to that of prior art 2, which is much less than that of prior art 1.

[0212] The present invention also provides an enterprise attribution system for industrial nodes, including:

[0213] Capture module, preprocessing module, attribution module, output module,

[0214] The capture module is used to capture patent text and send it to the preprocessing module;

[0215] The preprocessing module is used to preprocess the patent text, and send the patent text whose similarity with the target industrial chain exceeds the first threshold after preprocessing to the attribution module;

[0216] The attribution module is used to attribute the applicant enterprise of the patent text to the corresponding industrial node in the target industrial chain, and send the attribution result to the output module;

[0217] The output module is used to output the attribution results to the technicians;

[0218] Each module is programmed or configured to execute the steps of the above-mentioned enterprise attribution method for an industrial node.

[0219] The techniques, shapes, and structural parts not described in detail in the present invention are all well-known techniques.

[0220] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for enterprise attribution of industrial nodes, characterized in that: The following steps are involved: S1, build the target industrial chain; S2, after preprocessing the captured patent texts, the patent texts whose similarity with the target industry chain is below the first threshold are eliminated, and the patent texts whose similarity with the target industry chain exceeds the first threshold are sent to the trained attribution module; S3, the attribution module attributes the applicant enterprise of the patent text to the corresponding industrial node in the target industrial chain; Training the attribution module also includes the following steps: Step 1: Build a sample industrial chain; Step 2: After preprocessing the captured patent texts, the patent texts whose similarity with the sample industry chain is below the first threshold are eliminated, and the m patent texts whose similarity with the sample industry chain exceeds the first threshold are retained, which are recorded as m sample patents; Step 3, based on m sample patents, construct a first training set, a second training set and a third training set; wherein the first training set includes m sample patents and the industry nodes to which each sample patent belongs, the second training set includes m sample patents and g industry nodes related to each sample patent, wherein m is a positive integer and g is an integer greater than 1, and the third training set includes m sample patents; Step 4: the untrained attribution module includes a fine-tuning model, and the fine-tuning model is trained and optimized using the first training set to obtain a first model; Step 5, in the attribution module, a reward model is obtained based on the first model copy, and the reward model is trained and optimized using the second training set to obtain a second model; Step 6, using the second model and the third training set to train the first model and optimize it into a third model, at which point the attribution module training is completed; The specific steps of obtaining m sample patents in step 2 are the same as those in S2; In step 3, constructing the first training set also includes steps 31 and 32: Step 31, the sample patent data of the jth sample patent among m sample patents is recorded as P(j), 1≤j≤m, and j is an integer, and Y P(j) represents the first similarity score set between the sample patent data P(j) and each industrial node in the sample industrial chain, Y P(j) , max is the first similarity score set Y P(j) The largest first similarity score among the samples, P(j) and Y P(j) , max The corresponding industrial nodes are used as samples in the initial training set; Step 32, manually review the correctness of the samples in the initial training set, and record the reviewed samples as the first training set TR1; Constructing the second training set also includes steps 31′ to 32′: Step 31′, first use the bidirectional maximum matching method based on the second word segmentation dictionary D U The current sample patent data and each industry node are segmented, and then the jieba word segmentation is used based on the Xinhua Dictionary to segment the current sample patent data and each industry node. The obtained word segmentation is represented as one-hot and then added to the third dictionary D dict ; Step 32′, using the Word2Vec model to obtain the word feature vectors of the current sample patent data and the word feature vectors of each industry node; Step 33′, based on the word feature vectors of the current sample patent data and each industry node, calculate the second similarity score between the current sample patent data and each industry node and arrange them in descending order, and then use the industry nodes corresponding to the first g second similarity scores and the current sample patent data as samples of the second training set; Step 32' also includes the following sub-steps: Step 321′, set the context window size to o, and sequentially use each word in the current sample patent data as the central word, and obtain the context word set corresponding to each central word: Price(in t )={in t-o ,...,In t-q ,...,In t-1 ,In t+1 ...,In t+o } Among them, w t represents the tth word in the current sample patent data, t and o are both positive integers greater than 1, Cen(w t ) represents the participle w t The set of context words when it is the central word, w t-o represents the (to)th word in the current sample patent data, w t+o represents the (t+o)th word in the current sample patent data, w t-q represents the (tq)th word segment in the current sample patent data, 1≤q≤o and q is an integer. When tq=0, Price(in t )={in t-q+1 ,...,In t-1 ,In t+1 ...,In t+o }; Step 322′, using the Word2Vec model based on the third dictionary D dict The index position of each central word found in the search, the index position of the context word set, and the index position of the word segments contained in each industry node are used to obtain the corresponding central word feature vector, context word set feature vector, and each industry node feature vector; Note the current sample patent data center word w t The central word feature vector is Z(w t ), the central word w t The feature vector of the corresponding context word set is {Z(w t-o ),...,Z(w t-q ),...,Z(w t-1 ),Z(w t+1 )...,Z(w t+o )}, industrial node U in sample industrial chain U a The eigenvector of Z(U a ), Z(U a ) is the industrial node U a The sum of all word segmentation feature vectors in industry node U a The word segmentation of is obtained by step 31′; Step 323′, obtain the characteristic vector Z[P(j)] of the current sample patent data P(j): Z[P(j)]=Σ t Z(W t ), use cosine similarity to calculate the second similarity score between the feature vector Z[P(j)] and the feature vectors of each industrial node in the sample industrial chain U; ​​take the first g industrial nodes when the second similarity score is arranged in descending order, and use them together with the current sample patent data as the samples in the second training set TR2, where g is a positive integer less than the number of nodes in the sample industrial chain; In step 4, the following sub-steps are also included: Step 41, by respectively adding a first dimension reduction matrix A1 initialized by Gaussian distribution and a first dimension increase matrix B1 initialized by zero next to the attention network structure in an unsupervised LLM large language model, a fine-tuning model is obtained, and the attention network parameters in the unsupervised LLM large language model are recorded as θ0; Step 42, after converting the first sample in the first training set TR1 into the first prompt text, after training the fine-tuning model, calculate the first objective function objective1 after training using the current sample: The content of the first prompt text is {"instruction":"To which industry node does the knowledge involved in the following sample patent data belong?","input":"sample patent data","target":"industry node name"}; Among them, M represents the text content corresponding to the current target, x represents the concatenation of the text content in the current instruction and input, |M| represents the length of the text content corresponding to the current target, and M f Indicates the fth word in the text content corresponding to the current target, M <f Indicates that the index position of the current target in the text content is in M f All the words before; p LLM (M f |x,M <f ) means that only the LLM large language model in the current fine-tuning model is used to predict the probability of the fth word based on the first (f-1) words in the text content corresponding to the current target; p(M f |x,M <f ) means using the current fine-tuned model to predict the probability of the fth word based on the first (f-1) words in the text content corresponding to the current target; Step 43, take the first dimension reduction matrix copy A1′ and the first dimension increase matrix copy B1′ corresponding to the maximum value of the first objective function objective1 gradient rise, to iteratively update the corresponding first dimension reduction matrix A1 and the first dimension increase matrix B1 in the fine-tuning model, and at the same time update the attention network parameters in the LLM large language model in the fine-tuning model to obtain a new fine-tuning model: θ′=θ+A1′·B1′, where θ′ represents the updated attention network parameters, θ represents the attention network parameters before this update, and A1′·B1′ represents the dot product of A1′ and B1′; Step 44, after converting the next sample in the first training set TR1 into the first prompt text, the new fine-tuning model is trained, and steps 42 to 43 are repeated until the last sample in the first training set TR1 is used to train and update the fine-tuning model, and the fine-tuning model obtained at this time is recorded as the first model; In step 5, the following sub-steps are also included: Step 51, after removing the sfotmax function of the output layer in the first model copy, a reward model is obtained, and the reward model outputs a scalar, i.e., a score, according to the input; Step 52, after converting each sample in the second training set TR2 into the second prompt text, the text contents of instruction, input and target are concatenated as input to train the reward model, g = 2, and the reward model outputs the corresponding score pair, i.e., r(e, y1) and r(e, y2), The content of the second prompt text is {"instruction":"Which industry nodes does the knowledge involved in the following sample patent data belong to?", "input":"sample patent data", "target":["industry node name 1", "industry node name 2"]}, where the industry nodes in ["industry node 1", "industry node name 2"] are arranged in descending order according to the second similarity scores calculated in step 323'. e represents the concatenation of the text contents of instruction and input in the current second prompt text, y1 represents the text content of industry node name 1 in the current target, y2 represents the text content of industry node name 2 in the current target, r(e,y1) represents the score corresponding to y1 obtained based on e using the current reward model, and r(e,y2) represents the score corresponding to y2 obtained based on e using the current reward model; Step 53, after each Q1 sample is used to train the reward model, the second objective function objective2 is calculated once, and the reward model is optimized in the direction of the gradient descent of the second objective function objective2: Among them, (e, y1, y2)~TR2 represent the samples (e, y1, y2) in the second training set TR2, represents the expected value of the current sample (e, y1, y2), σ[·] represents the sigmoid function, 0<Q1≤m, and Q1 is a positive integer; Step 54, the reward model after repeating steps 52 to 53 for R1 training rounds is recorded as the second model. One training round means training the reward model once using all samples in the second training set TR2; In step 6, the following sub-steps are also included: Step 61, after converting each sample in the third training set TR3 into a third prompt text, the text contents of instruction and input in the third prompt text are concatenated as input to train the first model, and the first model outputs the corresponding answer, that is, the name of an industry node; The content of the third prompt text is {"instruction":"To which industry node does the knowledge involved in the following sample patent data belong?", "input":"sample patent data"}; Step 62, after each Q2 samples are used to train the first model, the current sample and the corresponding answer are converted into the first prompt text, and the text contents of instruction, input and target in the current first prompt text are concatenated as the input of the second model, and the output r(c,M) corresponding to the second model is obtained, wherein c represents the concatenation of the text contents of instruction and input in the current first prompt text, M represents the text content of target in the current first prompt text, i.e., the name of the industry node, and r(c,M) represents the score corresponding to M obtained based on c using the current second model, 0<Q2≤m, and Q2 is a positive integer; Calculate the third objective function objective3 once, and optimize the first model in the direction of the gradient ascent of the third objective function objective3: Among them, β is a constant parameter, (c)~TR3 means that the text content of the input in c is a sample in the third training set TR3, represents the expected value under the current c, π SFT (M|c) represents the probability that the initial first model obtains output M when the input is c, π RL (M|c) represents the probability that the current first model obtains the output M when the input is c, and the initial first model refers to the first model obtained in step 44; Step 63, after repeating steps 61 to 62 for R2 training rounds, the first model is recorded as the third model. One training round means training the first model once using all samples in the third training set TR3, and the attribution module training is completed at this time.

2. The enterprise attribution method of an industrial node according to claim 1, characterized in that: S2 specifically includes the following sub-steps: S21, after crawling the patent text, it is deduplicated and cleaned to obtain patent data; S22, constructing a first word segmentation dictionary of the target industrial chain, and segmenting the patent titles in the patent data according to the first word segmentation dictionary to obtain corresponding core contents; S23, after calculating the similarity between each core content and the target industrial chain, the patent data corresponding to the core content whose similarity with the target industrial chain is below a first threshold is eliminated.

3. The enterprise attribution method of an industrial node according to claim 2, characterized in that: The patent data obtained in S21 includes the patent name T of patent text i i and Instruction Manual Summary A i ; S22 also includes the following sub-steps: S221, after manually segmenting the names of each industry node, the segmentation results are added as standard segmentation words to the first segmentation dictionary D v In the example above, the part of speech of the standard participle is marked as noun; S222, based on the first word segmentation dictionary D v Patent Name T i After segmentation, retain the patent name T i The noun in is taken as the segmentation result; S223, remove patent name T i When the word frequency in the segmentation result is higher than the second threshold C2 and it is a noun of irrelevant vocabulary, the patent name T i The remaining nouns in the segmentation results are concatenated to obtain the corresponding core content S i : The second threshold Among them, δ is the first parameter, W max (Ti) indicates patent name T i The highest frequency of nouns in the segmentation results, W ave (Ti) indicates patent name T i The average frequency of all nouns in the segmentation results. The frequency of occurrence indicates the patent name T i The noun in the segmentation result appears in the corresponding instruction manual abstract A i The frequency of irrelevant words in the text includes: method, system, and device.

4. The enterprise attribution method of an industrial node according to claim 3, characterized in that: S23 also includes the following sub-steps: S231, calculating the first similarity score between each core content and each industry node, and obtaining a first similarity score set between each core content and each industry node: Y(S i )={Y(S i ,V1),...,Y(S i ,V k ),...,Y(S i ,V p )}; Among them, Y(S i ,V k ) indicates the core content S i With industry node V k The first similarity score between |S i ∩V k | indicates core contentS i With industry node V k The number of identical characters in the name, |S i ∪V k | indicates core contentS i With industry node V k The total number of characters in the name, Y(S i ) indicates the core content S i The first similarity score set with each industry node; target industry chain V = {V1, V2, ..., V k ,...,V p }, where V is the target industry chain, V1 represents the most upstream industry node in the target industry chain V, that is, the first industry node in the target industry chain V, and V p represents the most downstream industrial node in the target industrial chain V, that is, the pth industrial node in the target industrial chain V, and also the last industrial node, V k represents the kth industrial node in the target industrial chain V, 1≤k≤p, and both k and p are positive integers; S232, record Y(S i ) max is the first similarity score set Y(S i ), when Y(S i ) max When the first threshold C1 is below, the core content S is removed. i The corresponding patent data, on the contrary, will be related to the core content S i The corresponding patent data is sent to the trained attribution module; S3 also includes the following contents: the attribution module outputs the industry node corresponding to the current patent data. At the same time, the attribution module retrieves the patent text corresponding to the current patent data in the local database, retrieves the information of the applicant company from the patent text, and then binds the applicant company to the industry node currently output by the attribution module and outputs it to the technical personnel.

5. An enterprise attribution system for industrial nodes, characterized in that: include: Capture module, preprocessing module, attribution module, output module, The capture module is used to capture patent text and send it to the preprocessing module; The preprocessing module is used to preprocess the patent text, and send the patent text whose similarity with the target industrial chain exceeds the first threshold after preprocessing to the attribution module; The attribution module is used to attribute the applicant enterprise of the patent text to the corresponding industrial node in the target industrial chain, and send the attribution result to the output module; The output module is used to output the attribution results to the technicians; Each module is programmed or configured to execute the steps of the enterprise attribution method for an industrial node as described in any one of claims 1-4.

Images